Scholarlink

Background <p>Application of machine learning (ML) in neuroimaging is often constrained by the difficulty of assembling, sharing, and utilizing large, high-quality datasets. Synthetic data may address limitations in traditional data-sharing by enabling privacy-preserving image generation at scale. We evaluated whether: (1) a denoising diffusion probabilistic model (DDPM) could generate realistic synthetic lateral cervical spine radiographs, and (2) whether synthetic data could support transfer learning for anatomical landmark localization in low-data settings.</p> Methods <p>We trained a DDPM on 4,963 radiographs from the Cervical Spine X-ray Atlas (CSXA) to generate synthetic images. Blinded expert validation involved six neuroradiologists and two spine fellowship-trained neurosurgeons reviewing 50 randomly selected image quartets, each containing one real and three synthetic images. Experts attempted to identify the real scan and rated the realism of each image on a four-point Likert scale. To assess privacy preservation and potential memorization, we performed a real-synthetic nearest-neighbor search using vision transformer image embeddings ranked by cosine similarity. We then generated pseudolandmarks for our curated synthetic dataset and trained a landmark-localization backbone. External evaluation was performed on the Cervical Lateral X-ray 34-point (CLX-34) dataset using a directly matched 22-landmark task, comparing fine-tuning from the synthetic backbone and <i>de novo</i> training on increasing amounts of CLX-34 data.</p> Results <p>Experts correctly identified the real image in 29.0% of trials, with low inter-rater agreement (Fleiss’ κ = 0.061). Mean realism ratings did not differ significantly between real and synthetic images, and no visually explicit memorization was identified among the 100 most similar real-synthetic pairs. After manual quality control, the final released synthetic dataset comprised 20,063 radiographs. On the external CLX-34 test set (<i>n</i> = 374), fine-tuning from the synthetic backbone outperformed <i>de novo</i> training at 6 of 7 matched training sizes. The advantage was greatest in low-data settings, and the fine-tuned model achieved the best overall performance at <i>n</i> = 700 (mean MSE 4.37 vs. 7.42).</p> Conclusions <p>We present synthetic cervical spine radiographs that are statistically indistinguishable from real scans in blinded expert review and that provide useful initialization for landmark localization when external labeled data are limited.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Synthetic cervical spine radiographs: expert validation and transfer learning for low-data settings

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月11日

查看全文
  • Austin A. Barr,
  • Brij S. Karmur,
  • Anthony J. Winder,
  • Eddie Guo,
  • John T. Lysack,
  • James N. Scott,
  • William F. Morrish,
  • Muneer Eesa,
  • Morgan Willson,
  • David W. Cadotte,
  • Michael M. H. Yang,
  • Ian Y. M. Chan,
  • Sanju Lama,
  • Garnette R. Sutherland

摘要

Background

Application of machine learning (ML) in neuroimaging is often constrained by the difficulty of assembling, sharing, and utilizing large, high-quality datasets. Synthetic data may address limitations in traditional data-sharing by enabling privacy-preserving image generation at scale. We evaluated whether: (1) a denoising diffusion probabilistic model (DDPM) could generate realistic synthetic lateral cervical spine radiographs, and (2) whether synthetic data could support transfer learning for anatomical landmark localization in low-data settings.

Methods

We trained a DDPM on 4,963 radiographs from the Cervical Spine X-ray Atlas (CSXA) to generate synthetic images. Blinded expert validation involved six neuroradiologists and two spine fellowship-trained neurosurgeons reviewing 50 randomly selected image quartets, each containing one real and three synthetic images. Experts attempted to identify the real scan and rated the realism of each image on a four-point Likert scale. To assess privacy preservation and potential memorization, we performed a real-synthetic nearest-neighbor search using vision transformer image embeddings ranked by cosine similarity. We then generated pseudolandmarks for our curated synthetic dataset and trained a landmark-localization backbone. External evaluation was performed on the Cervical Lateral X-ray 34-point (CLX-34) dataset using a directly matched 22-landmark task, comparing fine-tuning from the synthetic backbone and de novo training on increasing amounts of CLX-34 data.

Results

Experts correctly identified the real image in 29.0% of trials, with low inter-rater agreement (Fleiss’ κ = 0.061). Mean realism ratings did not differ significantly between real and synthetic images, and no visually explicit memorization was identified among the 100 most similar real-synthetic pairs. After manual quality control, the final released synthetic dataset comprised 20,063 radiographs. On the external CLX-34 test set (n = 374), fine-tuning from the synthetic backbone outperformed de novo training at 6 of 7 matched training sizes. The advantage was greatest in low-data settings, and the fine-tuned model achieved the best overall performance at n = 700 (mean MSE 4.37 vs. 7.42).

Conclusions

We present synthetic cervical spine radiographs that are statistically indistinguishable from real scans in blinded expert review and that provide useful initialization for landmark localization when external labeled data are limited.